Optimizing Compiler for a CELL Processor

Total Page:16

File Type:pdf, Size:1020Kb

Optimizing Compiler for a CELL Processor Optimizing Compiler for a CELL Processor Alexandre E. Eichenberger† , Kathryn O’Brien†, Kevin O’Brien†, Peng Wu†, Tong Chen†, Peter H. Oden†, Daniel A. Prener†, Janice C. Shepherd† , Byoungro So†, Zehra Sura†, Amy Wang‡, Tao Zhang, Peng Zhao‡, and Michael Gschwind† †IBM T.J. Watson Research Center ‡IBM Toronto Laboratory College of Computing Yorktown Heights, New York, USA. Markham, Ontario, Canada. Georgia Tech, USA. ABSTRACT first generation CELL processor includes a 64-bit multi- Developed for multimedia and game applications, as well as threaded Power Processor Element (PPE) with two lev- other numerically intensive workloads, the CELL processor els of globally-coherent cache. It supports multiple op- provides support both for highly parallel codes, which have erating systems including Linux. For additional perfor- high computation and memory requirements, and for scalar mance, a CELL processor includes eight Synergistic Pro- codes, which require fast response time and a full-featured cessor Elements (SPEs). Each SPE consists of a new pro- programming environment. This first generation CELL pro- cessor designed for streaming workloads, a local memory, cessor implements on a single chip a Power Architecture pro- and a globally-coherent DMA engine. Computations are cessor with two levels of cache, and eight attached stream- performed by 128-bit wide Single Instruction Multiple Data ing processors with their own local memories and globally (SIMD) functional units. An integrated high bandwidth bus coherent DMA engines. In addition to processor-level par- glues together the nine processors and their ports to external allelism, each processing element has a Single Instruction memory and IO. Multiple Data (SIMD) unit that can process from 2 double In this paper, we present our compiler approach to sup- precision floating points up to 16 bytes per instruction. port the heterogeneous parallelism found in the CELL ar- This paper describes, in the context of a research pro- chitecture, which includes multiple, heterogeneous processor totype, several compiler techniques that aim at automat- elements and SIMD units on all processing elements. The ically generating high quality codes over a wide range of proposed approach is implemented as a research prototype heterogeneous parallelism available on the CELL proces- in IBM’s XL product compiler code base and currently sup- sor. Techniques include compiler-supported branch predic- ports the C and Fortran languages. tion, compiler-assisted instruction fetch, generation of scalar Our first contribution is a set of compiler techniques that codes on SIMD units, automatic generation of SIMD codes, provide high levels of performance for the SPE processors. and data and code partitioning across the multiple proces- To achieve high rates of computation at moderate costs in sor elements in the system. Results indicate that significant power and area, functionality that is traditionally handled in speedup can be achieved with a high level of support from hardware has been partially offloaded to the compiler, such the compiler. as memory realignment and branch prediction. We provide techniques to address these new demands in the compiler. Our second contribution is to automatically generate 1. INTRODUCTION SIMD codes that fully utilize the functional units of the The increasing importance of multimedia, game applica- SPEs as well as the VMX unit found on the PPE. The pro- tions, and other numerically intensive workloads has gener- posed approach minimizes the overhead due to misaligned ated an upsurge in novel computer architectures tailored for data streams and is tailored to handle many of the code such applications. Such applications include highly parallel structures found in multimedia and gaming applications. codes, such as image processing or game physics, which have Our third contribution is to enhance the programmabil- high computation and memory requirements. They also in- ity of the CELL processor by parallelizing a single source clude scalar codes, such as networking or game artificial in- program across the PPE and 8 SPEs. Key to this is our ap- telligence, for which fast response time and a full-featured proach of presenting the user with a single shared memory programming environment are paramount. image, effected through compiler mediated partitioning of Developed with such applications in mind, the CELL pro- code and data and the automatic orchestration of any data cessor provides both flexibility and high performance. This movement implied by this partitioning. We report an average speedup factor of 1.3 for the proposed SPE optimization techniques. When they are combined with SIMD and parallelization compiler tech- Permission to make digital or hard copies of all or part of this work for niques, we achieve average speedup factors of, respectively, personal or classroom use is granted without fee provided that copies are 9.9 and 7.1, on suitable benchmarks. Although not inte- not made or distributed for profit or commercial advantage and that copies grated yet, the latter two techniques will be cumulative as bear this notice and the full citation on the first page. To copy otherwise, to they address distinct sources of parallelism. republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 2005 IEEE X-XXXXX-XX-X/XX/XX ...$5.00. a) CELL Processor b) Synergistic Processing Element (SPE) 1 SPE SPE SPE SPE SPE SPE SPE SPE Even Pipe Odd Pipe Dual−Issue Floating/ Branch Instruction Fixed Memory Logic Point Permute Element Interconnect Bus, 96B/cycle Register File Instr. Buffer 128 x 16B 3.5 x 32 instr. 8 byte/cycle 2 in each dir. 16 byte/cycle Local Store in one dir. 256KB 32 byte/cycle Single Ported in one dir. 3 branch: 1,2 128 byte/cycle branch hint: 1,2 Power Processing Element (PPE) To External Mem To External IO in one dir. DMA instr fetch: 2 dma request: 3 Figure 1: The implementation of a first-generation CELL Processor This paper is organized as follows. We describe the CELL Instructions Pipe Lat. processor in Section 2 and our programming model in Sec- arithmetic, logical, compare, select even 2 tion 3. We present our SPE optimization techniques in Sec- shift, rotate, byte sum/diff/avg even 4 tion 4 and our automatic SIMD code generation techniques float even 6 in Section 5. We discuss our techniques for parallelization 16-bit integer multiply-accumulate even 7 and partitioning of single source programs among the PPE 128-bit shift/rotate, shuffle, estimate odd 4 and its 8 SPEs in Section 6. We report performance results load, store, channel odd 6 in Section 7 and conclude in Section 9. branch odd 1-18 2. CELL ARCHITECTURE Figure 2: Latencies and pipe assignment for SPE. The implementation of a first-generation CELL proces- sor [1] includes a Power Architecture processor and 8 at- An SPE can dispatch up to two instructions per cycle tached processor elements connected by an internal, high- to seven execution units that are organized into even and bandwidth Element Interconnect Bus (EIB). Figure 1a odd instruction pipes. Instructions are issued inorder and shows the organization of the CELL elements and the key routed to their corresponding even/odd pipe. Independent bandwidths between them. instructions are detected by the issue logic hardware and are The Power Processor Element (PPE) consists of a 64-bit, dual-issued provided they satisfy the following code-layout multi-threaded Power Architecture processor with two lev- conditions: the first instruction must come from an even els of on-chip cache. The cache preserves global coherence word address and use the even pipe, and the second in- across the system. The processor also supports IBM’s Vector struction must come from an odd word address and use the Multimedia eXtensions (VMX) [2] to accelerate multimedia odd pipe. When this condition is not satisfied, the two in- applications using its VMX SIMD units. structions are simply executed sequentially. The instruction A major source of compute power is provided by the eight latencies and their pipe assignments are shown in Table 2. on-chip Synergistic Processor Elements (SPEs) [3, 4]. An The SPE’s 256K-byte local memory supports fully- SPE consists of a new processor, designed to accelerate me- pipelined 16-byte accesses for memory instructions and 128- dia and streaming workloads, its local non-coherent mem- byte accesses for instruction fetch and DMA transfers. Be- ory, and its globally-coherent DMA engine. Key units and cause the memory is single ported, instruction fetches, bandwidths are shown in Figure 1b. DMA, and memory instructions compete for the same port. Nearly all instructions provided by the SPE operate in Instruction fetches occur during idle memory cycles, and up a SIMD fashion on 128 bits of data representing either 2 to 3.5 fetches may be buffered in the 32-instruction fetch 64-bit double floats or long integers, 4 32-bit single float buffers to better tolerate bursty peak memory usages. To or integers, 8 16-bit shorts, or 16 8-bit bytes. Instructions further avoid instruction starvation, an explicit instruction may source up to three 128-bit operands and produce one can be used to initiate an inline instruction fetch. 128-bit result. The unified register file has 128 registers and The branches are assumed non-taken by the SPE hard- supports 6 read and 2 write per cycle. ware but the architecture allows for a branch hint instruction The memory instructions also access 128 bits of data, with to override the default branch prediction policy. In addi- the additional constraint that the accessed data must reside tion, the branch hint instruction prefetches up to 32 instruc- at addresses that are multiples of 16 bytes. Namely, the tions starting from the branch target, so that a correctly- lower 4 bits of the load/store byte addresses are simply ig- hinted taken branch incurs no penalty.
Recommended publications
  • The LLVM Instruction Set and Compilation Strategy
    The LLVM Instruction Set and Compilation Strategy Chris Lattner Vikram Adve University of Illinois at Urbana-Champaign lattner,vadve ¡ @cs.uiuc.edu Abstract This document introduces the LLVM compiler infrastructure and instruction set, a simple approach that enables sophisticated code transformations at link time, runtime, and in the field. It is a pragmatic approach to compilation, interfering with programmers and tools as little as possible, while still retaining extensive high-level information from source-level compilers for later stages of an application’s lifetime. We describe the LLVM instruction set, the design of the LLVM system, and some of its key components. 1 Introduction Modern programming languages and software practices aim to support more reliable, flexible, and powerful software applications, increase programmer productivity, and provide higher level semantic information to the compiler. Un- fortunately, traditional approaches to compilation either fail to extract sufficient performance from the program (by not using interprocedural analysis or profile information) or interfere with the build process substantially (by requiring build scripts to be modified for either profiling or interprocedural optimization). Furthermore, they do not support optimization either at runtime or after an application has been installed at an end-user’s site, when the most relevant information about actual usage patterns would be available. The LLVM Compilation Strategy is designed to enable effective multi-stage optimization (at compile-time, link-time, runtime, and offline) and more effective profile-driven optimization, and to do so without changes to the traditional build process or programmer intervention. LLVM (Low Level Virtual Machine) is a compilation strategy that uses a low-level virtual instruction set with rich type information as a common code representation for all phases of compilation.
    [Show full text]
  • Analysis of Program Optimization Possibilities and Further Development
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Elsevier - Publisher Connector Theoretical Computer Science 90 (1991) 17-36 17 Elsevier Analysis of program optimization possibilities and further development I.V. Pottosin Institute of Informatics Systems, Siberian Division of the USSR Acad. Sci., 630090 Novosibirsk, USSR 1. Introduction Program optimization has appeared in the framework of program compilation and includes special techniques and methods used in compiler construction to obtain a rather efficient object code. These techniques and methods constituted in the past and constitute now an essential part of the so called optimizing compilers whose goal is to produce an object code in run time, saving such computer resources as CPU time and memory. For contemporary supercomputers, the requirement of the proper use of hardware peculiarities is added. In our opinion, the existence of a sufficiently large number of optimizing compilers with real possibilities of producing “good” object code has evidently proved to be practically significant for program optimization. The methods and approaches that have accumulated in program optimization research seem to be no lesser valuable, since they may be successfully used in the general techniques of program construc- tion, i.e. in program synthesis, program transformation and at different steps of program development. At the same time, there has been criticism on program optimization and even doubts about its necessity. On the one hand, this viewpoint is based on blind faith in the new possibilities bf computers which will be so powerful that any necessity to worry about program efficiency will disappear.
    [Show full text]
  • Register Allocation and Method Inlining
    Combining Offline and Online Optimizations: Register Allocation and Method Inlining Hiroshi Yamauchi and Jan Vitek Department of Computer Sciences, Purdue University, {yamauchi,jv}@cs.purdue.edu Abstract. Fast dynamic compilers trade code quality for short com- pilation time in order to balance application performance and startup time. This paper investigates the interplay of two of the most effective optimizations, register allocation and method inlining for such compil- ers. We present a bytecode representation which supports offline global register allocation, is suitable for fast code generation and verification, and yet is backward compatible with standard Java bytecode. 1 Introduction Programming environments which support dynamic loading of platform-indep- endent code must provide supports for efficient execution and find a good balance between responsiveness (shorter delays due to compilation) and performance (optimized compilation). Thus, most commercial Java virtual machines (JVM) include several execution engines. Typically, there is an interpreter or a fast com- piler for initial executions of all code, and a profile-guided optimizing compiler for performance-critical code. Improving the quality of the code of a fast compiler has the following bene- fits. It raises the performance of short-running and medium length applications which exit before the expensive optimizing compiler fully kicks in. It also ben- efits long-running applications with improved startup performance and respon- siveness (due to less eager optimizing compilation). One way to achieve this is to shift some of the burden to an offline compiler. The main question is what optimizations are profitable when performed offline and are either guaranteed to be safe or can be easily validated by a fast compiler.
    [Show full text]
  • Generalizing Loop-Invariant Code Motion in a Real-World Compiler
    Imperial College London Department of Computing Generalizing loop-invariant code motion in a real-world compiler Author: Supervisor: Paul Colea Fabio Luporini Co-supervisor: Prof. Paul H. J. Kelly MEng Computing Individual Project June 2015 Abstract Motivated by the perpetual goal of automatically generating efficient code from high-level programming abstractions, compiler optimization has developed into an area of intense research. Apart from general-purpose transformations which are applicable to all or most programs, many highly domain-specific optimizations have also been developed. In this project, we extend such a domain-specific compiler optimization, initially described and implemented in the context of finite element analysis, to one that is suitable for arbitrary applications. Our optimization is a generalization of loop-invariant code motion, a technique which moves invariant statements out of program loops. The novelty of the transformation is due to its ability to avoid more redundant recomputation than normal code motion, at the cost of additional storage space. This project provides a theoretical description of the above technique which is fit for general programs, together with an implementation in LLVM, one of the most successful open-source compiler frameworks. We introduce a simple heuristic-driven profitability model which manages to successfully safeguard against potential performance regressions, at the cost of missing some speedup opportunities. We evaluate the functional correctness of our implementation using the comprehensive LLVM test suite, passing all of its 497 whole program tests. The results of our performance evaluation using the same set of tests reveal that generalized code motion is applicable to many programs, but that consistent performance gains depend on an accurate cost model.
    [Show full text]
  • Efficient Symbolic Analysis for Optimizing Compilers*
    Efficient Symbolic Analysis for Optimizing Compilers? Robert A. van Engelen Dept. of Computer Science, Florida State University, Tallahassee, FL 32306-4530 [email protected] Abstract. Because most of the execution time of a program is typically spend in loops, loop optimization is the main target of optimizing and re- structuring compilers. An accurate determination of induction variables and dependencies in loops is of paramount importance to many loop opti- mization and parallelization techniques, such as generalized loop strength reduction, loop parallelization by induction variable substitution, and loop-invariant expression elimination. In this paper we present a new method for induction variable recognition. Existing methods are either ad-hoc and not powerful enough to recognize some types of induction variables, or existing methods are powerful but not safe. The most pow- erful method known is the symbolic differencing method as demonstrated by the Parafrase-2 compiler on parallelizing the Perfect Benchmarks(R). However, symbolic differencing is inherently unsafe and a compiler that uses this method may produce incorrectly transformed programs without issuing a warning. In contrast, our method is safe, simpler to implement in a compiler, better adaptable for controlling loop transformations, and recognizes a larger class of induction variables. 1 Introduction It is well known that the optimization and parallelization of scientific applica- tions by restructuring compilers requires extensive analysis of induction vari- ables and dependencies
    [Show full text]
  • The Effect of Code Expanding Optimizations on Instruction Cache Design
    May 1991 UILU-EN G-91-2227 CRHC-91-17 Center for Reliable and High-Performance Computing THE EFFECT OF CODE EXPANDING OPTIMIZATIONS ON INSTRUCTION CACHE DESIGN William Y. Chen Pohua P. Chang Thomas M. Conte Wen-mei W. Hwu Coordinated Science Laboratory College of Engineering UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN Approved for Public Release. Distribution Unlimited. UNCLASSIFIED____________ SECURITY ¿LASSIPlOvriON OF t h is PAGE REPORT DOCUMENTATION PAGE 1a. REPORT SECURITY CLASSIFICATION 1b. RESTRICTIVE MARKINGS Unclassified None 2a. SECURITY CLASSIFICATION AUTHORITY 3 DISTRIBUTION/AVAILABILITY OF REPORT 2b. DECLASSIFICATION/DOWNGRADING SCHEDULE Approved for public release; distribution unlimited 4. PERFORMING ORGANIZATION REPORT NUMBER(S) 5. MONITORING ORGANIZATION REPORT NUMBER(S) UILU-ENG-91-2227 CRHC-91-17 6a. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION Coordinated Science Lab (If applicable) NCR, NSF, AMD, NASA University of Illinois N/A 6c ADDRESS (G'ty, Staff, and ZIP Code) 7b. ADDRESS (Oty, Staff, and ZIP Code) 1101 W. Springfield Avenue Dayton, OH 45420 Urbana, IL 61801 Washington DC 20550 Langley VA 20200 8a. NAME OF FUNDING/SPONSORING 8b. OFFICE SYMBOL ORGANIZATION 9. PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER 7a (If applicable) N00014-91-J-1283 NASA NAG 1-613 8c. ADDRESS (City, State, and ZIP Cod*) 10. SOURCE OF FUNDING NUMBERS PROGRAM PROJECT t a sk WORK UNIT 7b ELEMENT NO. NO. NO. ACCESSION NO. The Effect of Code Expanding Optimizations on Instruction Cache Design 12. PERSONAL AUTHOR(S) Chen, William, Pohua Chang, Thomas Conte and Wen-Mei Hwu 13a. TYPE OF REPORT 13b. TIME COVERED 14. OATE OF REPORT (Year, Month, Day) Jl5.
    [Show full text]
  • Eliminating Scope and Selection Restrictions in Compiler Optimizations
    ELIMINATING SCOPE AND SELECTION RESTRICTIONS IN COMPILER OPTIMIZATION SPYRIDON TRIANTAFYLLIS ADISSERTATION PRESENTED TO THE FACULTY OF PRINCETON UNIVERSITY IN CANDIDACY FOR THE DEGREE OF DOCTOR OF PHILOSOPHY RECOMMENDED FOR ACCEPTANCE BY THE DEPARTMENT OF COMPUTER SCIENCE SEPTEMBER 2006 c Copyright by Spyridon Triantafyllis, 2006. All Rights Reserved Abstract To meet the challenges presented by the performance requirements of modern architectures, compilers have been augmented with a rich set of aggressive optimizing transformations. However, the overall compilation model within which these transformations operate has remained fundamentally unchanged. This model imposes restrictions on these transforma- tions’ application, limiting their effectiveness. First, procedure-based compilation limits code transformations within a single proce- dure’s boundaries, which may not present an ideal optimization scope. Although aggres- sive inlining and interprocedural optimization can alleviate this problem, code growth and compile time considerations limit their applicability. Second, by applying a uniform optimization process on all codes, compilers cannot meet the particular optimization needs of each code segment. Although the optimization process is tailored by heuristics that attempt to a priori judge the effect of each transfor- mation on final code quality, the unpredictability of modern optimization routines and the complexity of the target architectures severely limit the accuracy of such predictions. This thesis focuses on removing these restrictions through two novel compilation frame- work modifications, Procedure Boundary Elimination (PBE) and Optimization-Space Ex- ploration (OSE). PBE forms compilation units independent of the original procedures. This is achieved by unifying the entire application into a whole-program control-flow graph, allowing the compiler to repartition this graph into free-form regions, making analysis and optimization routines able to operate on these generalized compilation units.
    [Show full text]
  • Compiler-Based Code-Improvement Techniques
    Compiler-Based Code-Improvement Techniques KEITH D. COOPER, KATHRYN S. MCKINLEY, and LINDA TORCZON Since the earliest days of compilation, code quality has been recognized as an important problem [18]. A rich literature has developed around the issue of improving code quality. This paper surveys one part of that literature: code transformations intended to improve the running time of programs on uniprocessor machines. This paper emphasizes transformations intended to improve code quality rather than analysis methods. We describe analytical techniques and specific data-flow problems to the extent that they are necessary to understand the transformations. Other papers provide excellent summaries of the various sub-fields of program analysis. The paper is structured around a simple taxonomy that classifies transformations based on how they change the code. The taxonomy is populated with example transformations drawn from the literature. Each transformation is described at a depth that facilitates broad understanding; detailed references are provided for deeper study of individual transformations. The taxonomy provides the reader with a framework for thinking about code-improving transformations. It also serves as an organizing principle for the paper. Copyright 1998, all rights reserved. You may copy this article for your personal use in Comp 512. Further reproduction or distribution requires written permission from the authors. 1INTRODUCTION This paper presents an overview of compiler-based methods for improving the run-time behavior of programs — often mislabeled code optimization. These techniques have a long history in the literature. For example, Backus makes it quite clear that code quality was a major concern to the implementors of the first Fortran compilers [18].
    [Show full text]
  • An ECMA-55 Minimal BASIC Compiler for X86-64 Linux®
    Computers 2014, 3, 69-116; doi:10.3390/computers3030069 OPEN ACCESS computers ISSN 2073-431X www.mdpi.com/journal/computers Article An ECMA-55 Minimal BASIC Compiler for x86-64 Linux® John Gatewood Ham Burapha University, Faculty of Informatics, 169 Bangsaen Road, Tambon Saensuk, Amphur Muang, Changwat Chonburi 20131, Thailand; E-mail: [email protected] Received: 24 July 2014; in revised form: 17 September 2014 / Accepted: 1 October 2014 / Published: 1 October 2014 Abstract: This paper describes a new non-optimizing compiler for the ECMA-55 Minimal BASIC language that generates x86-64 assembler code for use on the x86-64 Linux® [1] 3.x platform. The compiler was implemented in C99 and the generated assembly language is in the AT&T style and is for the GNU assembler. The generated code is stand-alone and does not require any shared libraries to run, since it makes system calls to the Linux® kernel directly. The floating point math uses the Single Instruction Multiple Data (SIMD) instructions and the compiler fully implements all of the floating point exception handling required by the ECMA-55 standard. This compiler is designed to be small, simple, and easy to understand for people who want to study a compiler that actually implements full error checking on floating point on x86-64 CPUs even if those people have little programming experience. The generated assembly code is also designed to be simple to read. Keywords: BASIC; compiler; AMD64; INTEL64; EM64T; x86-64; assembly 1. Introduction The Beginner’s All-purpose Symbolic Instruction Code (BASIC) language was invented by John G.
    [Show full text]
  • The Jalapeño Dynamic Optimizing Compiler for Javatm
    TM The Jalap eno ~ Dynamic Optimizing Compiler for Java Michael G. Burke Jong-Deok Choi Stephen Fink David Grove Michael Hind Vivek Sarkar Mauricio J. Serrano V. C. Sreedhar Harini Srinivasan John Whaley IBM Thomas J. Watson Research Center P.O. Box 704, Yorktown Heights, NY 10598 Abstract p erformance and scalabilityofJava applicatio ns on SMP server machines. Some previous high p erformance imple- The Jalap eno ~ Dynamic Optimizing Compiler is a key com- mentations of Java e.g., [9, 25, 22, 18] have relied on static 1 p onent of the Jalap eno ~ Virtual Machine, a new Java Vir- compilation, and have therefore disallowed certain features tual Machine JVM designed to supp ort ecient and scal- such as dynamic class loading. In contrast, the Jalap eno ~ able execution of Java applications on SMP server machines. JVM supp orts all features of Java [29], and the Jalap eno ~ This pap er describ es the design of the Jalap eno ~ Optimizing 2 Optimizing Compiler is a fully integrated dynamic compiler Compiler, and the implementation results that wehave ob- in the Jalap eno ~ JVM. tained thus far. To the b est of our knowledge, this is the rst The Jalap eno ~ pro ject was initiated in Decemb er 1997 at dynamic optimizing compiler for Java that is b eing used in a the IBM T. J. Watson Research Center and is still work-in- JVM with a compile-only approach to program execution. progress. This pap er describ es the design of the Jalap eno ~ Optimizing Compiler and the implementation results that 1 Intro duction wehave obtained thus far.
    [Show full text]
  • Register Allocation Deconstructed
    Register Allocation Deconstructed David Ryan Koes Seth Copen Goldstein Carnegie Mellon University Carnegie Mellon University Pittsburgh, PA Pittsburgh, PA [email protected] [email protected] Abstract assignment. The coalescing component of register allocation Register allocation is a fundamental part of any optimiz- attempts to eliminate existing move instructions by allocat- ing compiler. Effectively managing the limited register re- ing the operands of the move instruction to identical loca- sources of the constrained architectures commonly found in tions. The spilling component attempts to minimize the im- embedded systems is essential in order to maximize code pact of accessing variables in memory. The move insertion quality. In this paper we deconstruct the register allocation component attempts to insert move instructions, splitting the problem into distinct components: coalescing, spilling, move live ranges of variables, with the goal of achieving a net im- insertion, and assignment. Using an optimal register alloca- provement in code quality by improving the results of the tion framework, we empirically evaluate the importance of other components. The assignment component assigns those each of the components, the impact of component integra- program variables that aren’t in memory to specific hard- tion, and the effectiveness of existing heuristics. We evalu- ware registers. Existing register allocators take different ap- ate code quality both in terms of code performance and code proaches in solving these various components. For example, size and consider four distinct instruction set architectures: the central focus of a graph-coloring based register allocator ARM, Thumb, x86, and x86-64. The results of our investiga- is to solve the assignment problem, while the spilling and tion reveal general principles for register allocation design.
    [Show full text]
  • Design and Evaluation of Register Allocation on Gpus
    Design and Evaluation of Register Allocation on GPUs A Thesis Presented by Charu Kalra to The Department of Electrical and Computer Engineering in partial fulfillment of the requirements for the degree of Master of Science in Electrical and Computer Engineering Northeastern University Boston, Massachusetts November 2015 ⃝c Copyright 2015 by Charu Kalra All Rights Reserved i Abstract The rapid adoption of Graphics Processing Unit (GPU) computing has led to the develop- ment of workloads which can leverage the massive parallelism offered by GPUs. The use of GPUs is no longer limited to graphics, but has been extended to support compute-intensive applications from various scientific and commercial domains. Significant effort and expertise are required to optimize applications targeted for a specific GPU architecture. One of the keys to improving the performance of the applications is to utilize the available hardware resources efficiently. The archi- tectural complexity of the GPU makes it challenging to optimize performance at both the source level and the compiler level. In this thesis, we develop a transformation pass (i.e., a register allocator) that optimizes both vector and scalar register usage on GPUs. We introduce an open-source compiler framework, Multi2C, which converts OpenCL kernel into AMD Southern Islands binaries. The register alloca- tor is integrated as a part of the Multi2C optimization framework. Register allocation for each class of registers is run as a separate pass in the compiler. We also highlight the challenges faced when performing register allocation targeting a GPU, including issues with thread divergence and proper handling of contiguous register series. We evaluate the performance of our allocator for applications taken from the AMD APP SDK.
    [Show full text]